Original Link: http://www.anandtech.com/show/1109
Computex Half-a-World Away - Part 1: Intel Roadmaps Revealedby Anand Lal Shimpi on June 2, 2003 2:23 AM EST
- Posted in
- Trade Shows
This article first appeared on the AnandTech Newsletter. Click Here to Sign Up.
Normally, every June the AnandTech Staff gets geared up for the 14 - 18 hours of plane travel to take us half way around the world to Taipei, Taiwan for the annual Computex tradeshow. Because of the threat of SARS, the manufacturers and event partners have delayed this year's show until the week of September 22nd instead of its usual period the first week of June. The hope is that SARS will be more under control by this Fall, but even despite this some manufacturers are already stating that they won't be attending the show even in the Fall.
With Computex being my personal favorite show out of all of those we attend, and with the high quality of information usually given out at the show, it would definitely be a shame to potentially miss out on that this year - thus the idea of doing a remote Computex was born.
The idea is simple; instead of flying to Taiwan to see all of the products and meet with the manufacturers, we met via email and got the products sent to us via carrier in order to bring you some of the sort of coverage you could expect from the show. We've split the coverage into three parts, the first focusing on Intel's roadmaps (as usual, the motherboard guys are great for giving us information on Intel's plans), the second focusing on Small Form Factor solutions and finishing up with coverage on what's new from the motherboard guys.
There's no way to duplicate the greatness of the Computex show without actually having one, but we'll try our best to give you all the information you're looking for sans show.
Only 3 new Desktop CPUs this year
In 2002 Intel released an incredible 7 CPUs for the desktop, but believe it or not, you're only going to see three more CPUs between now and the end of the year.
As we mentioned in our last Intel Roadmap update, the 3.2GHz Pentium 4 is next up to bat and it will be the last 0.13-micron Northwood processor before Prescott. The 3.2GHz Northwood Pentium 4 will continue to be the fastest thing from Intel even through Q3 2003, before Prescott's release in Q4.
We've described Prescott in great detail in previous articles, so here's some information on it if you're a bit rusty on the specs and benefits of Intel's first 90nm CPU:
Things don't heat up until Q4, when we see the long awaited introduction of the Prescott core. Right now Intel is the unequivocal performance leader on the desktop, but what do users of the fastest CPUs today have to look forward to? Prescott of course. We've been preaching for months that Prescott would be the next CPU to upgrade to (much like we suggested waiting on migrating to the Pentium 4 until the Northwood core was released), and that time is almost upon us.
The Prescott core will be introduced at two clock speeds - 3.2GHz and 3.4GHz. The 3.2GHz part will obviously overlap with the forthcoming 3.2GHz Northwood part, while the 3.4GHz speed will be a Prescott-only core. The other thing to keep in mind is that both of these CPUs are still Socket-478 parts, you won't see the first Socket-775 CPUs until Q2 2004, which we will talk about next.
The 3.4GHz Prescott will close off 2003, but in Q1 '04 we'll see another speed bump to 3.60GHz. This 3.60GHz processor will be the one to stay away from, because it is still a Socket-478 CPU. In the following quarter, Q2 '04, Intel will introduce a Socket-775 LGA version of the 3.60GHz Prescott processor as well as 775-pin versions of the 3.2/3.4GHz Prescott processors.
There is a chance that we will see a 3.80GHz Prescott core in Q2 '04 as well, however that will largely depend on how much competition Intel is faced with. If you recall, the major difference between 2002 and 2003 is that for the majority of 2002, AMD was a much more potent competitor. We will have to wait and see how the launch of the Athlon 64 goes before determining how big of a threat AMD will be going forward on the desktop side. One thing is for sure though; it's going to take a lot more than 64-bit support for AMD to pull sales away from Intel on the desktop side.
Grantsdale & Socket-775
In our review of the 865 line of chipsets we provided a brief introduction to the Canterwood/Springdale successor known as Grantsdale. Grantsdale will accompany the 3.20 - 3.60GHz Socket-775 Prescott CPUs launched in Q2 '04, and will feature the following specs:
- 800MHz FSB support (initially, moving up to 1.06/1.2GHz eventually)
- Dual DDR400/333 SDRAM
- Dual DDR2-400/533 SDRAM (probably reserved for the higher end Grantsdale chipsets, and obviously depending on market availability of DDR2 SDRAM)
- Optional integrated graphics (Grantsdale-G) using a brand new integrated graphics core
- One PCI Express x16 slot for graphics
- ICH6 with support for 4 independent Serial ATA channels and PCI Express x1 slots
The first thing to take away from these specs is that Intel is expecting DDR2 SDRAM to be available in decent enough quantities by Q2 '04 in order to incorporate it into their roadmap. We've seen a clear departure from the RDRAM dominated roadmaps of Intel's past and the trend continues as we look towards 2004.
The presence of Dual DDR2-533 on the roadmap indicates a desired transition to the 1.06GHz FSB, as Intel has been stressing synchronous FSB/memory bus operation for quite some time now. It's interesting that there's no mention of CSA support for Grantsdale chipsets, but it could just be that Grantsdale will be accompanied with an unannounced derivative of CSA. We have little reason to believe that CSA would be changed at all for Grantsdale, it could just be that Intel doesn't see the need to mention the bus much like they don't mention the Hub Link revisions each chipset features (Hub Link connects the MCH/North Bridge to the ICH/South Bridge).
Intel does have a backup strategy, should DDR2 support not be as prevalent as desired, which is to continue to offer DDR support. Luckily, from a bandwidth standpoint, even if Grantsdale launches with no more than DDR400 support, Intel won't pay any performance penalties. The true benefit of enabling DDR2 support is to push for higher speed memory buses beyond 400MHz, which will become necessary as FSB speeds reach 1.06GHz and beyond.
In the unlikely event that DDR2 fails to catch on, Intel could always migrate to quad-channel RDRAM; at PC1066 speeds, a quad-channel RDRAM solution would still be able to deliver as much bandwidth as a 1.06GHz FSB would require at 8.5GB/s. Unfortunately once you move to four 16-bit RDRAM channels, you begin to lose some of the low pincount advantage of RDRAM as a solution - granted that we're still talking about lower pincounts than a 128-bit DDR interface, but the advantage isn't as great anymore. It's interesting how things have changed; at one point, SDR/DDR SDRAM were Intel's fall-back solutions to RDRAM, but now only in the worst case scenario does RDRAM have a place on Intel's memory roadmap.
Graphics will change slightly with Grantsdale; from an integrated standpoint, we'll see the 3rd generation "Extreme Graphics" core in Grantsdale-G. We would assume that Intel would take this opportunity to introduce DX9 class hardware, considering it will be introduced almost two years after ATI put forth their first DX9 GPU. Unfortunately, implementing 24-bit FP pixel pipelines in addition to 32-bit FP vertex pipelines in addition to all of the programmable logic that was missing from the 2nd generation "Extreme Graphics" core would make Grantsdale-G a very large chip.
The high transistor count of a full DX9 part leads us to believe that Intel may favor a partial DX9 implementation ranging from either a core with only partial precision support (16-bit FP precision) or continuing their trend of not including full hardware support in their graphics core. Considering that Longhorn (Microsoft's 3D UI) will include both DX7 and DX9 compatibility modes, Intel could get away with not including a full DX9 implementation in their Extreme Graphics 3 core. Only time will tell but we're not overly confident in Intel doing something incredible with Extreme Graphics 3. Remember that it's in Intel's best interests to place as much stress on owning a fast CPU as possible; offloading work onto a powerful GPU definitely doesn't vibe with their intentions.
The other unique thing about Grantsdale is the introduction of a PCI Express x16 slot instead of an AGP slot for discrete graphics. ATI and NVIDIA have already announced support for PCI Express interfaces on their graphics cards, initially using an AGP-PCI Express bridge, but eventually they will offer native PCI Express cards. We would expect some Grantsdale boards to ship with AGP support as you can't really expect users that buy $500 video cards today to toss them in order to move to Grantsdale.
PCI Express will also be seen on Grantsdale boards in the form of x1 slots, but the usefulness of those slots will be determined by the availability of PCI Express devices.
Finally we see that Intel has increased the number of Serial ATA channels from 2 to 4 with the introduction of a new ICH - ICH6/ICH6R. The Grantsdale chipsets will still feature the same number of USB 2.0 ports as Springdale/Canterwood (8).
The market has never seen such a powerful launch in the mobile sector as Intel's Centrino; shortly after the mobile platform's launch, we are already seeing exciting solutions from all of the major manufacturers. A good part of this article itself was written on a Centrino based mobile platform, there's very little doubt that Centrino (or more appropriately, Pentium-M and the Odem chipset) is the perfect choice for all non-desktop replacement laptops.
Unforuntately, Centrino solutions are not cheap these days (neither are most laptops for that matter) and the question that is often brought up is - when will today's Centrino solutions be obsolete?
Unlike the original Pentium 4, the next Pentium-M won't be such a drastic departure from the current design. With that said, the improvements you can expect from Dothan (Banias'successor) are not insignificant. As we've mentioned countless times before, Dothan will be the first 90nm processor for the mobile market. Interestingly enough, Intel plans on introducing Dothan around the same time as Prescott - meaning we'll be seeing two 90nm introductions in Q4 of this year. There have been a number of times where Intel has debuted a smaller manufacturing process on the mobile side first, before transitioning their desktop parts; with both Prescott and Dothan due out in the 4th quarter, which one comes first is a toss up.
As we mentioned in our Pentium-M (Banias) microarchitecture article, Intel's Israel design team essentially built-in clock speed walls into the processor's architecture. For more information on why this was done be sure to read our review of the CPU, but as a quick recap, in order to minimize the amount of power used Intel's designers had to design the Pentium-M processor in such a way that the CPU would not be able to ramp in clock frequency all that well. The only hope for increasing clock frequency significantly on the Pentium-M, without significant architecture enhancements, would be to manufacture the processor on smaller, faster transistors. In the case of Dothan, using Intel's 90nm process will give the processor some extra frequency headroom. How much headroom?
Currently, the Banias processor is available in speeds of up to 1.60GHz, however in Q3 of this year Intel will introduce a 1.70GHz Pentium-M processor. This won't be the processor to upgrade to, and we'll explain why in a moment. In the same quarter Intel will extend the Low-Volt and Ultra Low-Volt Pentium-M lines to include 1.20GHz and 1.0GHz parts respectively. The reason you won't want to consider any of these processors is because in the next quarter, Q4 '03, Intel will finally introduce Dothan - the 90nm Pentium-M processor with a full 2MB on-die L2 cache.
Dothan will still feature the same 400MHz FSB (it won't really need a 533MHz or faster FSB until the processor starts to ramp well above 2GHz) and will feature some minor architecture enhancements but its biggest feature will be the 2MB on-die L2 cache. Dothan will be introduced at 1.80GHz and we're assuming based on Intel's roadmap, in relatively low quantities. If you're planning a Centrino notebook upgrade in the 2nd half of this year, then you may want to try and wait for Dothan at 1.80GHz. If you must have something sooner, then rest assured that you won't be buying anything obsolete for this year as the processor will top out at 1.80GHz.
In the first half of 2004 we will begin to see Dothan ramp up with 1.90GHz and 2.0GHz versions. We will also see the first Low-Volt and Ultra Low-Volt Dothan processors in the first half, weighing in at 1.30GHz and 1.0GHz respectively. In Q2 '04 we will see 1.10GHz and faster ULV Dothan processors.
Along with Dothan's introduction, Intel will also be bringing a new Wireless LAN chip to market - the Intel PRO/Wireless 2100A. The 2100A will support both 802.11a and 802.11b wireless LAN standards. As you will remember from our Centrino coverage, in order for a manufacturer to use the "Centrino" name they must use Intel's Pentium-M processor, 855 chipset, and PRO/Wireless LAN; unfortunately, Intel currently has no 802.11a solution so if a manufacturer wants to have 802.11a support in their Centrino solution they're out of luck. Instead, manufacturers have been using other 802.11a solutions and resorted to calling their notebooks Pentium-M based solutions, which adds to an already confusing market.
The introduction of an Intel 802.11a solution will definitely help, but the high price Intel is charging for their 802.11x chips will guarantee that manufacturers will continue to make both Centrino and simple Pentium-M based solutions. Intel is planning on pricing the PRO/Wireless 2100A at an astounding $65, that's almost 45% more than their PRO/Wireless 2100 is going for (in 1,000 unit quantities). Of course, Intel will have their usual "rebates" in place to encourage manufacturers to support the Centrino platform instead of using their own wireless solutions combined with a Pentium-M processor.
The beginning of 2004 will see the introduction of an 802.11g PRO/Wireless solution from Intel, although that could happen anytime during the 1st half of next year depending on finalization of the 802.11g spec.
Intel's 855GM chipset will receive a refresh to coincide with Dothan's launch with the 855GME. It is unclear what the enhancements will be and it's also interesting that there's no mention of an 'E' successor to the 855PM, could the improvements only be graphics related (the 855GM chipset features integrated graphics)?
Mobile Pentium 4 - Not Dead
Despite all of the attention that has been paid to Centrino and the Pentium-M processor, Intel still sees a viable market for the Pentium 4 on the mobile side. The desktop replacement and larger thin-and-light segments will both continue to be serviced by the Mobile Pentium 4 processor.
In Q3, Intel will introduce a 3.06GHz Mobile Pentium 4 without Hyper-Threading support for the mobile market. It won't be until Q4 that we see Hyper-Threading on mobile CPUs, with the introduction of a 3.20GHz Mobile Pentium 4. Intel will then release a HT enabled 3.06GHz Mobile Pentium 4; as to why Intel doesn't just wait or enable HT initially with the first 3.06GHz mobile chip, your guess is as good as ours. It seems detrimental to the market to release two 3.06GHz parts, one with and one without Hyper-Threading within such a close time period of one another.
The other thing to keep in mind is that the mobile platforms will not receive the new 800MHz FSB, nor will they receive derivatives of the new chipsets (865/875) anytime soon - thus the 3.2GHz Mobile Pentium 4 is still a 533MHz part. For battery life, the 3.2GHz Mobile Pentium 4 is probably going to be one of the worst choices to make - but for the target market of desktop replacements, battery life isn't a major concern.
In the Q1 '04, Intel will introduce the Mobile Prescott processor at 2.66, 2.80, 3.06, 3.20 and 3.46GHz, all using the 533MHz FSB and all equipped with Hyper-Threading support. In the following quarter, Intel will ramp clock speeds above 3.46GHz for the Mobile Prescott CPUs. The Mobile Prescott should have all of the same features as the desktop CPUs, including the large 1MB L2 cache (not large by Dothan standards, but by all other standards); the Mobile Prescott is, of course, built on Intel's 90nm manufacturing process.
Servers & Workstations Get Confusing
Intel's strategy in the server and workstation market has become somewhat cluttered thanks to the plethora of CPUs you'll find available to suit the various segments; not only do we have the desktop Prescott processor being targeted at entry-level servers/workstations, but then there's the regular Xeon DP, the 1MB L2 Xeon DP, the Xeon MP and Itanium. Within those families there are also 800MHz, 533MHz and 667MHz FSB versions, and don't always assume that the faster, more expensive processors can take advantage of the higher frequency FSBs. Confused yet? Let's get to the roadmaps
We've already proven that Canterwood (875P) has virtually no use on the desktop from a performance standpoint; Intel's 865PE chipset is just about as fast and the motherboard manufacturers are working on tweaking their BIOSes to make the solutions even faster and more competitive with the 875P. Remember that the motherboard guys want to sell 865PE platforms more than 875P since they target a lower price point and thus can meet their volume requirements, the margins on 875P boards aren't high enough for the motherboard guys to get too excited. But remember from our review of the 875P chipset that one of the features is ECC memory support - why include ECC support? For the server/workstation markets of course, this is where 875P is truly made for.
Intel's roadmaps show the 875P chipset being used in all uniprocessor (UP) workstations and a derivative of the chipset called Canterwood-ES used in all UP servers. The server derivative of Canterwood most likely features some increased validation for peace of mind in the server environment.
Both the Canterwood-ES and 875P chipsets are supposed to be paired up with Intel's Prescott CPU, and will feature 800MHz FSB support. Interestingly enough, this is the only mention of 800MHz FSB support that exists on Intel's server/workstation roadmap. Instead, all other segments will feature either a 533 or 667MHz FSB going forward. As far as why, we can't tell, it could be tied to increased validation at 800MHz, unwillingness for many enterprise customers to adopt DDR400, or a number of other theories - regardless, don't expect any 800MHz FSB solutions to penetrate beyond the entry-level workstation/server markets.
In Q3 of this year Intel will finally introduce their "Opteron competitor" for the dual processor (DP) markets - a 3.06GHz Xeon with a 1MB on-die L2 cache. As you can guess, this is a 533MHz FSB part, and will exist alongside a 512KB L2 part. In the following quarter, Intel will introduce the Nocona processor, which is basically a Prescott (90nm) derived core for the enterprise market.
Unlike the desktop Prescott processor however, Nocona will only support the 533MHz FSB at launch (much like mobile Prescott); the processor will be introduced at 3.20 and 3.46GHz. The beginning of 2004 will allow Nocona to enjoy a slightly faster FSB of 667MHz, to help it reach speeds of 3.50, 3.60GHz and beyond. The lack of any 800MHz FSB support is puzzling, since it would come in handy in these markets - arguably more than on the desktop.
Continuing on into the first half of 2004, Nocona will continue to be offered in both 533 and 667MHz FSB flavors.
Nocona will be aimed mainly at the UP/DP server and workstation markets and will eventually be paired with the Lindenhurst chipset - a successor to the E7501 with 667MHz FSB support. On the workstation side, the Tumwater chipset will take over as the successor to the E7505.
From a performance standpoint, things make even less sense once we start looking at Intel's plans for the Xeon MP line for 4-way configurations. Intel's Xeon MPs will continue to use a 400MHz FSB through the first half of 2004, even as the processors reach speeds as high as 2.60GHz. Intel's reasoning behind this is mostly to allow for a single upgrade path without forcing enterprise customers to upgrade their motherboards as frequently as we have to on the desktop side, but it's clear that the performance of the processors will be limited by FSB bandwidth. Remember that Intel's MP solutions rely on a shared bus protocol, meaning that in a 4-way Xeon MP server all four processors must share the same 400MHz FSB. This essentially quadruples the FSB bandwidth requirements of the server, and the Xeon MP happens to be the processor that will have the least amount of FSB bandwidth out of all of Intel's enterprise CPUs.
Potomac will be the first 90nm Xeon MP processor, but we won't see that until the second half of 2004. Considering Intel's 0.13-micron Xeon MPs may have up to 4MB of on-die L3 cache, we can only assume that Potomac will have anywhere from 6 - 8MB of cache on-die.
To round out Intel's IA-32 enterprise roadmap, we have the Pentium-M and Dothan processors being used in 1U Low Power blade servers. At the last Intel Developer Forum we showed off some demonstrations of the Pentium-M (Banias) processor being used in a server environment, well you can expect that to continue with Dothan.
No article on Intel's enterprise roadmap would be complete without a mention of Itanium. Intel's Itanium 2 continues to be a presence on the server/workstation roadmaps as expected, but what's interesting is the introduction of Deerfield which will happen sometime in Q3 of this year.
As you will remember from our IDF coverage, Deerfield is going to be the entry-level Itanium processor. Intel is positioning Deerfield as a Low-Voltage Itanium 2, and it will be made available at speeds as high as 1.0GHz with a 1.5MB on-die L3 cache. It isn't clear how Deerfield will differ architecturally from the regular Itanium 2 core, but Intel's roadmaps do mention that the part will be DP capable only. Intel could feasibly make Deerfield as capable as the regular Itanium 2, but only limit it to 2-way configurations, which would succeed in limiting competition within the Itanium family.
What's most intriguing about the Low-Voltage Itanium 2 (Deerfield) is its price tag. The 1.0GHz part will sell for under $800 in 1,000 unit quantities, compared to the $2,000+ that the current Itanium 2 1.0GHz goes for.
Every Computex we get a good deal of information about Intel's forthcoming plans, with this year's show moved to the Fall we weren't about to let a good update on Intel's roadmap slip us by.
Intel has been coming on strong in the past couple of years, the ball has been in AMD's court for quite some time now - Athlon 64 and Opteron both need to do well in order for AMD to regain lost ground.
What we're hearing from motherboard manufacturers and the server guys is that Opteron is doing quite well in the same areas that the Athlon MP has sold well in (HPC, clusters, etc ) however they have yet to see much expansion above and beyond where AMD used to be.
The idea of a single processor Opteron workstation (using nForce3 Pro for example) continues to be an interesting one, but the closer we get to AMD's September launch date for Athlon 64, the less interest there will be in 1P Opteron machines. ASUS had better get those nForce3 Pro boards out and onto the market soon in order to capture the opportunity.
Athlon 64 motherboards have been ready for quite some time, and the motherboard makers are growing impatient waiting for AMD's launch. Luckily, it seems as if the 865-line of chipsets will be increasing demand for products from our friends in Taiwan. The one to benefit from the delays the Athlon 64 has seen seems to be NVIDIA, as the nForce2 is doing quite well in the Socket-A market.
More what-you-would've seen from Computex coverage to come, so stay tuned